Skip to content

Response, New York. Noah Park: Check Point Patches Actively Exploited Management Server Path Traversal. Detection, Dallas. Ava Okello: Report Finds Nearly Half of Deployed Detections Need Attention Before SOCs Can Trust Them.

SOCtember

Always First. Fast SOC News.

Elastic Security Labs illustration of an AI-driven intrusion: a hooded operator and a robot figure at terminals linked to cloud servers marked with keys, with lines fanning out to cloud, database and workload icons that each raise an alert.
Image: Elastic (fair use). Header art from the Elastic Security Labs post "56 AWS API calls in 141 milliseconds: Inside an AI-run intrusion," published Nov. 13, 2026.

Detection

Elastic Says a CyberStrikeAI Server Read 56 EC2 User Data Scripts in 141 Milliseconds and Worked Eight Stolen AWS Identities Over 16 Days, Yet All 78 Alerts It Set Off Were Medium or Low Severity

The intrusion only made sense once the alerts were grouped by source IP, and a 14-day pause would have split it apart under any fixed correlation window. Elastic has added two high-severity bulk enumeration rules and lists the CloudTrail patterns to hunt for.

Ava Okello, DetectionLondon6 min read

LONDON - A server running the open-source CyberStrikeAI attack platform spent 16 days inside four Amazon Web Services accounts using stolen credentials, and every one of the 78 detections it triggered looked routine on its own, Elastic Security Labs said in an analysis that lays out why AI-assisted cloud intrusions are hard to see in a standard alert queue.

Between Sept. 30 and Oct. 16, the server at 43[.]240[.]29[.]11 used credentials for eight identities, five IAM users and three EC2 instance roles, in a familiar order: validate identities, pull secrets from Secrets Manager, read EC2 user data, decrypt an SSM parameter and send a remote command, wrote researchers Samir Bousseaden, Bryan Porras Blanch and Terrance DeJesus. "Every one of those is a call a legitimate role makes every day," they wrote.

CyberStrikeAI pairs large language model agents with offensive security tools. Elastic cited Will Thomas of Team Cymru, who counted 21 IPs running it in early 2026, all aimed at Fortinet edge devices, and said this was the first case it had seen of the same tooling turned on AWS. The company called the activity AI-assisted based on the tool attribution and machine-speed behavior in the logs, but said it cannot prove an LLM agent was making decisions and does not know how the attacker got the credentials. The three role sessions were named after EC2 instance IDs, which means the credentials came from instance profiles and were used from outside, for example after server-side request forgery or a host compromise.

Machine speed, then silence

On the first day, the server called GetCallerIdentity on four IAM users across three accounts within 51 minutes, three of those calls in the same second. An EC2 role called app-role then read a secret and, about six hours later, fired 56 requests for instance user data, the startup scripts that often carry hardcoded keys and database passwords, in 141 milliseconds, or about 2.5 milliseconds per call. On Oct. 2, a third role read an encrypted SSM SecureString parameter and issued an SSM SendCommand 29 seconds later; CloudTrail does not show what the command was or whether it succeeded.

Then the server went quiet for 14 days. On Oct. 16 it came back and revalidated credentials, including a user in a fourth account that had not appeared before. "That gap did more damage than the burst, because fixed correlation windows cannot span two weeks," the researchers wrote.

Elastic bar chart of 78 AWS alerts by API call, colored by detection rule, dominated by 58 DescribeInstanceAttribute user data alerts, with 11 GetCallerIdentity, 4 GetSecretValue and 5 others.
Image: Elastic (fair use).Figure 1 from the Elastic post, alerts by AWS API call and detection rule.

Six rules, no high-severity alert

Six Elastic rules fired and together covered every phase: EC2 user data retrieval (58 alerts), first-time GetCallerIdentity (11), first-time Secrets Manager access (4), GetSessionToken usage (3), SSM SendCommand by a rare user (1) and an SSM SecureString request with decryption (1). All were medium or low severity, and 65 of the 78 came on the first day. Only the shared source IP showed they belonged together.

Elastic's answer is two new high-severity ES|QL rules for the bulk behavior. AWS EC2 Bulk User Data Enumeration fires when one principal and source IP read user data from more than 20 distinct instances in 10 minutes. AWS STS GetCallerIdentity Bulk Account Enumeration fires when three or more distinct non-role identities call GetCallerIdentity from the same external IP and ASN in 10 minutes. Medium and low alerts are still worth correlating, the researchers said, but by source IP and with a flexible window. They also showed an Elastic Workflows proof of concept that deactivates keys, revokes sessions and quarantines the IAM user when a rule fires.

Elastic Workflow named Remediate AWS - Stolen Access Key running steps that deactivate access keys, revoke sessions and quarantine an IAM user.
Image: Elastic (fair use).An Elastic Workflows proof of concept for automated IAM containment, from the Elastic post.

What to hunt and fix

Elastic's CloudTrail hunting list: bursts of DescribeInstanceAttribute calls for userData from one role session; first-time GetCallerIdentity calls from one external IP across several accounts or users in the same minute; GetSessionToken or GetSecretValue from an instance role session whose source IP is neither the instance nor the organization's NAT; and GetParameter with decryption followed within a minute by SSM SendCommand from the same role. It also published the infrastructure: two more IPs, 43[.]240[.]29[.]28 and 43[.]240[.]29[.]33, served the same self-signed CyberStrikeAI certificate on port 8080, though none of the three was serving the interface as of November, so they are best used for retrospective hunting.

For exposed accounts, Elastic recommends rotating credentials and revoking sessions for every identity that touched the indicator IP; replacing the instance profile or instance for stolen role credentials, since they can be reissued while the host stays compromised; treating all secrets in user data as exposed and moving them to Secrets Manager or Parameter Store; reviewing SSM command history, including use of AWS-RunShellScript and AWS-RunPowerShellScript; checking non-production secrets for overlap with production; restricting ec2:DescribeInstanceAttribute for userData to the roles that need it; and hunting for how the instance credentials were taken, starting with exposed web applications, SSRF and IMDS.

About the Author

Ava Okello

Ava Okello covers detection engineering, EDR telemetry, and SOC hunting for SOCtember from London.

Related stories

Detection desk